NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips

Valerio Venceslai, Alberto Marchisio, Ihsen Alouani, Maurizio Martina, Muhammad Shafique

TL;DR

NeuroAttack is a novel cross-layer attack that bridges the gap between software-level adversarial noise and hardware-level fault injection. By inserting a "sneaky" hardware Trojan that is only triggered by a specific, nearly invisible input pattern, researchers demonstrated they could completely collapse the accuracy of Spiking Neural Networks (SNNs) and Deep Neural Networks (DNNs) using as few as four surgical bit-flips.

What Problem Does This Paper Address?

As Spiking Neural Networks (SNNs) move from academic curiosity to mainstream deployment in energy-efficient hardware (like Intel’s Loihi), their security becomes a critical concern. While traditional Deep Neural Networks (DNNs) are known to be vulnerable to adversarial examples (maliciously crafted inputs) and fault-injection attacks (physically flipping bits in memory), SNN security is less understood.

The authors identify a major security gap: most attacks are either purely software-based or require physical access to hardware. This paper introduces a "cross-layer" threat where a hardware vulnerability (a Trojan) is hidden in the circuit during manufacturing but remains dormant until it is "woken up" by a specific high-level input pattern. This allows a remote attacker to compromise a system's integrity without needing physical access to the device during operation.

Methods

The NeuroAttack framework consists of three main stages:

  1. Vulnerability Analysis: The researchers used a gradient-based search algorithm to identify the most "sensitive" bits in the network's weights. They found that while SNNs are resilient to random bit-flips, they are incredibly fragile when specific bits are targeted.
  2. Hardware Trojan Design: They designed a lightweight hardware Trojan consisting of a 2-way multiplexer and an inverter. When inactive, the circuit passes the original weight bit; when triggered, it flips the bit. To make this work for SNNs, they added a counter to track spike rates, as SNNs communicate via timing and frequency rather than static values.
  3. Trigger Generation (Algorithm 1): This is the "key" to the attack. The authors developed an optimization loop to craft a specific noise pattern (the trigger) that maximizes the activation of a single "target neuron." When this neuron's output exceeds a threshold, it activates the hardware Trojan. The goal was to make this noise pattern imperceptible to humans and ensure it never triggers accidentally on clean, "natural" data.

Datasets

The authors evaluated NeuroAttack using standard computer vision benchmarks:

  • MNIST: Used to test a Multilayer Perceptron (MLP) and the LeNet-5 architecture.
  • CIFAR-10: Used to test a more complex Convolutional Neural Network (CNN) to prove the attack scales to more difficult tasks.

Key Insights and Ideas

  • Surgical Precision vs. Brute Force: The paper highlights that SNNs have an "inherent resilience" to random errors. You might need to flip thousands of random bits to see a drop in accuracy. However, using their gradient search, the authors proved that flipping just 30 bits (out of millions) can reduce accuracy to near zero. On CIFAR-10, only 4 bits were needed to crash the model.
  • The Neuron as a Remote Control: The most innovative idea is using a specific neuron's activation as the Trojan's "on" switch. By optimizing an input pattern to over-activate one specific internal neuron, the attacker creates a remote-controlled trigger that bypasses traditional hardware security checks.
  • Stealth is Prioritized: The attack is designed to be "stealthy" in two ways: the hardware overhead is minimal (only a few hundred transistors), and the input trigger is mathematically optimized to be nearly invisible noise, ensuring the model behaves perfectly normally until the specific trigger is present.

Key Results

  • Effectiveness: On the MNIST dataset, the attack achieved a "false activation" rate of zero—meaning the Trojan never accidentally turned on when processing clean images—while successfully triggering on 100% of the malicious inputs.
  • Hardware Overhead: For a typical attack, the hardware cost is roughly 540 transistors. In the context of modern chips with billions of transistors, this is an incredibly small and difficult-to-detect footprint.
  • Cross-Domain Transfer: The authors successfully demonstrated that vulnerabilities identified in the DNN domain (continuous values) could be effectively transferred to the SNN domain (discrete spikes), proving that SNN conversion doesn't inherently protect against these types of attacks.

Limitations and Future Work

  • Supply Chain Assumption: The attack assumes the adversary has "grey-box" access—meaning they know the architecture and can insert logic into the hardware during the design or manufacturing phase (a "supply chain attack"). It does not address how to insert such a Trojan into an already-deployed, "clean" chip.
  • Trigger Specificity: While the trigger is stealthy, it is specific to the network's weights. If the weights are updated (e.g., via on-chip learning), the original trigger might no longer activate the target neuron.
  • Future Defense: The authors suggest that future work should focus on "hardware-software co-design" for security, perhaps by developing specialized monitors that detect unusual spikes in internal neuron activity that don't correlate with typical input features.

Summary updated

Abstract

Due to their proven efficiency, machine-learning systems are deployed in a wide range of complex real-life problems. More specifically, Spiking Neural Networks (SNNs) emerged as a promising solution to the accuracy, resource-utilization, and energy-efficiency challenges in machine-learning systems. While these systems are going mainstream, they have inherent security and reliability issues. In this paper, we propose NeuroAttack, a cross-layer attack that threatens the SNNs integrity by exploiting low-level reliability issues through a high-level attack. Particularly, we trigger a fault-injection based sneaky hardware backdoor through a carefully crafted adversarial input noise. Our results on Deep Neural Networks (DNNs) and SNNs show a serious integrity threat to state-of-the art machine-learning techniques.
BibTeX
@inproceedings{venceslai2020neuroattack,
  author = {Valerio Venceslai and Alberto Marchisio and Ihsen Alouani and Maurizio Martina and Muhammad Shafique},
  title = {{NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips}},
  booktitle = {IJCNN},
  year = {2020},
  url = {https://openreview.net/forum?id=FiVc0ke7HW_},
}

Loading related content...

Figure